How to make custom pointcloud data as inference input?

[YiChenCityU]

Hi, Congratulations. I want to test the inference code with the pointcloud as input. Could you provide some advices? Thanks very much.

[SimonGiebenhain]

Hi @YiChenCityU,
thanks for your interest.
For the "dummy_dataset" as well as our proposed test set we provide single view data already.

If you want to change some properties of the input for inference you can play araound with scripts.data_processing.generate_single_view_observations.
I used this script to generate the input. Per default it tries to do so for every subject in the test set. But you can simply specifiy what subject and expression you are interested in.

[YiChenCityU]

Thanks very much. What if I only have a point cloud captured from iphone, do I have to provide the expression of it?

[YiChenCityU]

This is the point cloud I used and the result was not similar to it. Do you have some suggestions? Ply files are below.
https://drive.google.com/file/d/1UYBbR-TkRtgSKJQbuNUnMu4dwdN1kx9a/view?usp=sharing https://drive.google.com/file/d/1A4EJbSUjuAfJ_k8FzmsimsSKBPi1QQ5k/view?usp=sharing

[SimonGiebenhain]

Hey, cool stuff.

The problem is very likely the coordinate system. NPHM only works if the input is in the expected coordinate system (FLAME coordinate system scaled by a factor of 4).

Therefore, you would first have to align the input point cloud with the FLAME coordinate system, e.g. a very simple approach would be a similarity transform from detected 3D landmarks to the landmarks of the FLAME template. Actually, you could also first fit FLAME and use the resulting Scale, Rotation, and Translation from the result. In that case, you can separate the head from the torso in the same way as in the preprocessing of NPHM. Having observations on the torso tends to confuse the inference optimization

Here is an example mesh from the dataset and one of the provided point clouds to show why the model fails:

[SimonGiebenhain]

Actually, the second Point Cloud aligns better, but is still noticeably off from the expected canonicalization.

[YiChenCityU]

I will try. Thanks very much.

[xvdp]

Ive been trying to unravel the description as well, I didnt get as far as yichen, It would be wonderful if you could provide a full test example ...
If you are concerned about identity, maybe take a pointcloud of a statue...

[nsarafianos]

Thank you so much @SimonGiebenhain for publishing the code and congrats for your great work!

Quick Q: I have a pointcloud in.obj format (lifted from a foreground RGB-D monocular image) that is transformed to be on the exact same space with FLAME as suggested above. How do you go about fitting NPHM to this particular pointcloud ?

I'm asking because the example provided uses existing identities (along with their expressions) from the dummy_data whereas I'm interested in preserving the identity of the pointcloud.

Thank you!

[Zvyozdo4ka]

@SimonGiebenhain
even with perfect alignment it did not resemble the identity

Original files are here.
https://drive.google.com/drive/folders/1cprPG_9AihL4HpYl0lOvZDz7kNbXv8kB?usp=sharing

[Zvyozdo4ka]

The problem is very likely the coordinate system. NPHM only works if the input is in the expected coordinate system (FLAME coordinate system scaled by a factor of 4).

How did you get FLAME models? What solution did you employ?

Therefore, you would first have to align the input point cloud with the FLAME coordinate system, e.g. a very simple approach would be a similarity transform from detected 3D landmarks to the landmarks of the FLAME template.

Do you have this code of alignment or did you use another method to align point cloud and flame?

Actually, you could also first fit FLAME and use the resulting Scale, Rotation, and Translation from the result. In that case, you can separate the head from the torso in the same way as in the preprocessing of NPHM. Having observations on the torso tends to confuse the inference optimization

Do you mean that fitting Flame to point cloud can give us the same NPHM output?

[Zvyozdo4ka]

Actually, the second Point Cloud aligns better, but is still noticeably off from the expected canonicalization.

in your work, did you align point cloud and flame manually or using an alignment algorithm?

[mwb3262716541]

Actually, the second Point Cloud aligns better, but is still noticeably off from the expected canonicalization.

in your work, did you align point cloud and flame manually or using an alignment algorithm?

Hi, I have seen many your questions and answers under this project. Now I meet the same problem on converting the point cloud scan to the input cloud, have you solved this problem?

[Zvyozdo4ka]

Actually, the second Point Cloud aligns better, but is still noticeably off from the expected canonicalization.

in your work, did you align point cloud and flame manually or using an alignment algorithm?

Hi, I have seen many your questions and answers under this project. Now I meet the same problem on converting the point cloud scan to the input cloud, have you solved this problem?

i did not manage to resolve it, and you can see they've never responded me back
I tried manual alignment of photogrammetry or point avatar generated point clouds, but outputs comprises a lot of artifacts, and can't be used. From 20-30 experiment to manual align, only one sometimes can look well.
And they do not mention it directly in paper, but they used point cloud reconstructed from Kinect device, which i do not have, so i gave up.

You may consider trying MonoNPHM project from the same author.
But unfortunately my outputs did not resemble identity

[mwb3262716541]

Actually, the second Point Cloud aligns better, but is still noticeably off from the expected canonicalization.

in your work, did you align point cloud and flame manually or using an alignment algorithm?

Hi, I have seen many your questions and answers under this project. Now I meet the same problem on converting the point cloud scan to the input cloud, have you solved this problem?

i did not manage to resolve it, and you can see they've never responded me back I tried manual alignment of photogrammetry or point avatar generated point clouds, but outputs comprises a lot of artifacts, and can't be used. From 20-30 experiment to manual align, only one sometimes can look well. And they do not mention it directly in paper, but they used point cloud reconstructed from Kinect device, which i do not have, so i gave up.

You may consider trying MonoNPHM project from the same author. But unfortunately my outputs did not resemble identity

thanks for your reply. It is a little difficult. I noticed DPHM project recently. They provide code to preprocess the kinect depth map and convert it to the input inference. I am trying to run it but still not success.
one more question, how do you manully align point cloud and flame?

[Zvyozdo4ka]

thanks for your reply. It is a little difficult. I noticed DPHM project recently. They provide code to preprocess the kinect depth map and convert it to the input inference. I am trying to run it but still not success. one more question, how do you manully align point cloud and flame?

just manually in Cloud Compare, rotated, scaled. Also you can try in Blender